Graph Convolution Network with Double Filter for Point Cloud Segmentation

Abstract

To solve the problem of information loss caused by point cloud segmentation using voxels. A method of transforming point cloud into graph data and using double filter graph convolution network for segmentation is proposed. The first filter is for point clouds to reduce the number of nodes in the graph. Considering the feature as a signal, the convolution is defined in the spectral domain using a Laplacian matrix, and the Chebyshev polynomial is used to reduce the computational complexity of the matrix decomposition. The second filter is a low-pass filter for the Chebyshev polynomial, which reduce the computation. Finally, the 2D data is detected using CNN to optimizes the segmented result. Experiments were performed on the ShapeNet dataset to demonstrate the efficiency of the method.